Predicting corticosteroid resistance and recurrence in keloids: a machine learning study integrating dermatoscopy and ultrasound.

Background: Keloids are characterized by high recurrence rates and variable responses to intralesional corticosteroid therapy. Objective and non-invasive approaches for predicting treatment response and recurrence remain limited. This study aimed to develop and validate machine learning models integrating dermatoscopic and high-frequency ultrasound features for individualized risk stratification in patients with keloids. Methods: In this retrospective cohort study, 174 patients with keloids from the internal cohort and an independent external validation cohort of 30 patients were included. Dermatoscopic vascular patterns and ultrasound-derived lesion characteristics were analyzed. Four machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), Naive Bayes (NB), and XGBoost, were developed to predict corticosteroid treatment response and recurrence risk. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, calibration analysis, and decision curve analysis. Results: For corticosteroid response prediction, the RF model achieved the highest discrimination performance (AUC = 0.816, 95% CI: 0.74-0.88), followed by SVM (AUC = 0.781), NB (AUC = 0.759), and XGBoost (AUC = 0.744). For recurrence prediction, the XGBoost model showed the highest predictive performance (AUC = 0.949, 95% CI: 0.91-0.98). External validation in an independent 30-patient cohort showed comparable discrimination (RF: AUC = 0.851; XGBoost: AUC = 0.855), providing preliminary evidence for model reproducibility. Calibration and decision curve analyses suggested potential clinical usefulness. Conclusions: Machine learning models integrating dermatoscopic and ultrasound features may provide non-invasive approaches for predicting corticosteroid response and recurrence in keloids. These findings support further multicenter validation before clinical implementation.